基於賽局方法為物聯網的可拆分之深度神經網路推論在邊緣運算中進行資源管理

Abstract

隨著各種物聯網 (IoT) 系統需求的快速增長，深度學習 (DL) 相關的應用也引起了廣泛關注。然而，在計算能力有限的物聯網系統上執行深度學習模型的推論 (inference) 非常具有挑戰性。與雲端計算相比，邊緣計算 (edge computing) 將資源部署在終端用戶附近以減少傳輸延遲並保留原始數據以減輕隱私洩漏的問題。但由於邊緣的資源還是有限的，用戶跟邊緣節點之間的關聯 (association) 決策、邊緣節點的資源分配、使用的深度學習模型參數選擇，以及設備配置等管理就變得至關重要。此外，3GPP 最近為下一代無線網絡討論的機器學習推論分割 (split-ML) 概念也為邊緣管理提供了更大的靈活性。在本文中，我們首先通過 Vickrey-Clarke-Groves 機制，在考慮物聯網設備偏好的同時研究準確率和每秒幀數 (FPS) 要求下的邊緣節點資源分配問題。然後，加入機器學習推論分割中的隱私問題和多視角檢測提供的偵測準確度增強，我們將研究擴大到物聯網設備與邊緣節點之間的關聯決策問題。我們使用賽局中形成聯盟 (coalition formation) 的演算法來啟發式地優化社會福利，並證明最終聯盟結構的收斂性和穩定性。模擬結果表明，這兩個問題中的設計旋鈕都是必不可少的，並且確實在不同的情景設置下為邊緣運算的管理帶來了更好的整體效用。

Deep learning (DL) applications have attracted significant attention with the rapidly growing demand for Internet of Things (IoT) systems. However, it is challenging to perform the inference of such resource-hungry DL models on the computationally limited IoT system. Compared to cloud computing, edge computing deploys resources near the end-users to reduce the transmission delay while retaining the raw data to mitigate privacy concerns. Since resources at the edge are limited, management like user association decisions, edge resources allocation, and device configuration with DL model parameter selection becomes essential. Moreover, the concept of split-ML recently discussed by 3GPP for the next-generation wireless networks also provides more flexibility for management at the edge. In this work, we first study the joint edge resources allocation problem under accuracy and FPS requirements while considering IoT devices' preferences by applying the Vickrey–Clarke–Groves-based method. Then, we enlarge the problem to make the association decision between IoT devices and the edge nodes, considering the privacy retain issue in split-ML and the enhancement provided by multi-view detection. We handle the association problem heuristically by a coalition formation-based algorithm and prove properties such as convergence and stability. The simulation results show that the design knobs in both problems are essential and indeed lead to better social welfare under different scenario settings.